netinet6: fix interface handling for loopback traffic

Currently, processing of IPv6 local traffic is partially broken:
link-local connection fails and global unicast connect() takes
3 seconds to complete.
This happens due to the combination of multiple factors.
IPv6 code passes original interface "origifp" when passing
traffic via loopack to retain the scope that is mandatory for the
correct hadling of link-local traffic. First problem is that the logic
of passing source interface is not working correcly for TCP connections,
resulting in passing "origifp" on the first 2 connection attempts and
lo0 on the subsequent ones. Second problem is that source address
validation logic skips its checks iff the source interface is loopback,
which doesn't cover "origifp" case.
More detailed description is available at https://reviews.freebsd.org/D35732

Fix the first problem by untangling&simplifying ifp/origifp logic.
Fix the second problem by switching source address validation check to
using M_LOOP mbuf flag instead of interface type.

PR: 265089
Reviewed by: ae, bz(previous version)
Differential Revision: https://reviews.freebsd.org/D35732
MFC after: 2 weeks